Services

PepDesigner

Structure-based Peptide Binder Designer

A platform that designs new binding peptides from a target protein structure and ranks the candidates with molecular dynamics.

Open PepDesigner

What it does

PepDesigner starts from a target structure and generates sequences that might bind, then puts them in order. It designs with BindCraft, filters on predicted metrics, and computes the binding free energy with molecular dynamics for the top candidates only. You use it to decide which candidates to synthesise first.

All of it is computational prediction. No experimental validation is included.

How you use it

The screen has two columns. You configure on the left and watch progress and results on the right.

1. Target and binding site - Enter one of: protein name, UniProt accession, PDB ID, or sequence (FASTA). If you entered a name, click the magnifier and pick from the candidates - You can also upload a structure or sequence file (.pdb .ent .fasta .fa .seq .txt) - If the structure has several chains, choose the target chain - Choose the domain of interest. For a membrane protein you can turn on Extracellular domain only - Enter hotspot residues, comma-separated. Leave it blank and they are found automatically (by PyMOL surface exposure)
2. Binding partner (optional) — turn it on if you need it and enter the partner along with its domain and hotspots
3. Peptide design — number to generate (5 by default), length range (65–150 by default), number of GPUs, whether to design per epitope
4. MD and binding free energy — decide how many of the top candidates go through MD (10 by default)
5. Closed-loop redesign — number of iterations (3 by default)
6. Run analysis — start it
7. Progress — per-step status, refreshed every 3 seconds. Stop job halts it
8. Results — read the ranking, check it in 3D, and download the report

When it stops halfway and asks you

If you give the target as a UniProt accession alone and specify no hotspots, partner or domain, the pipeline fetches the target information and then stops to ask which site to attack. The domain list comes with surface scores, evidence grades and estimated times. You either pick one here, or press Let PepDesigner choose (auto) and let the service probe 3 to 6 sites itself.

Pipeline steps

The remaining time shown on screen is a rough figure correct only to the order of magnitude. It is not a measurement.

What is implemented

Specifying the target — UniProt REST name search, direct accession or PDB ID, sequence input, file upload. An uploaded structure is read for its chain count and per-chain lengths, which are shown to you.

Choosing the site — InterPro domain list, extracellular domain filter based on membrane topology, manual hotspots, automatic PyMOL hotspots, interactive domain picker, automatic search mode.

Design — BindCraft (AF2 hallucination + ProteinMPNN + AF2 rescoring). Independent attempts are launched at once across several GPUs and whichever fills the target count first is used.

Filtering — six metrics.

Binding free energy — MD is run with GROMACS and the calculation done with gmx_MMPBSA. The top three per-residue contributions come with it.

Viewing results — ranking table, an in-browser 3D viewer (Cartoon, Stick, Sphere, Line, rotation), report.md · report.html · report.pdf, and five figures.

Job management — recent job list (10 at a time), reattaches to a running job after a refresh, selective deletion, and deleting a running job releases its GPU.

Accounts — email and password plus Google sign-in, an admin screen, per-account job isolation, and account deletion.

What you have to know

The pass thresholds quietly loosen. If no design passes the thresholds above, they are lowered by up to 8 steps and the filter is run again until something comes out (each step is i_pTM −0.05, pLDDT −0.05, i_pAE +2.0, ΔG +10.0, SC −0.05, RMSD +0.5). And yet the report always prints the original strict thresholds. Do not judge by the "passed the filter" marker alone — compare the actual metric values against the table above yourself.

The binding free energy must not be read as an absolute value. The screen says MM-PBSA, but it is really MM-GBSA (a single GB-OBC2 protocol), the default production MD is short at about 100 ps, and no entropy term is computed. Use it only for relative comparison between candidates. If you need a value for a paper, you have to run the MD for longer.

A hotspot is not the same as efficacy. The designability search only asks "can a binder be made here". It does not judge whether blocking that site is therapeutically meaningful.

ERROR is not a verdict on that site. It means the search could not run. Do not confuse it with FLATLINED (cannot be designed).

They are long for "peptides". The default length range is 65–150 residues. If you want short peptides you have to narrow the range yourself.

Sequences leave the server. When you give only a sequence and the structure is predicted, Boltz-2 fetches the MSA from an external ColabFold server. Target information lookups (UniProt, InterPro, STRING, AlphaFold DB) are external calls too. Bear this in mind if the sequence is confidential.

A receptor that is too large gets cut. Above 450 residues a window is cut around the hotspots and used instead. This is a GPU memory limit.

What it cannot do

There is no queue. Press run and it starts immediately. Launch several and they compete for GPUs — if no GPU is free, they wait.

You cannot use sequence mode from the screen. The path that builds candidates from a sequence and ranks them with docking, BLAST and a quick MD exists in the code, but the screen is fixed to de novo design alone. It is available from the CLI only.

You cannot choose the structure prediction engine. It is fixed to Boltz-2. An AlphaFold2 path exists but cannot be selected on screen, and does not actually run either because of server permissions.

You cannot choose the MD engine. The default path runs on gmx_MMPBSA regardless of what you ask for.

There is no RFdiffusion and no ESM. Design is the one BindCraft (AF2 + ProteinMPNN) route.

There is no immunogenicity assessment. T-cell and B-cell epitope prediction is planned only; there is no code.

There is no experimental validation. No value in the ranking table means binding has been confirmed. The job is not finished until the top few candidates have actually been synthesised and checked.

There is no payment or billing. Accounts are approved automatically as soon as you register.

Software availability

PepDesigner does not compute anything itself. At each step it calls outside programs and models — BindCraft designs, Boltz-2 predicts structures, GROMACS runs the dynamics. So if one of them is not prepared on the server, that step does not run.

Unlike AURORA, there is no environment check screen that stops you before you start. Press run and it starts immediately; if something is missing, the step that needs it fails when the run reaches it. The earlier steps are already done, so check the progress panel and the result folder to see how far it got.

What each step uses

The statuses below are as of 2026-09-29, checked against the shared runtime for the program and the model weights actually being present. Readiness can change, so if you are reading this long afterwards, confirm with an administrator.

Note that design and scoring are bundled into one environment. A single BindCraft environment holds AF2, ProteinMPNN and PyRosetta together. So if that environment breaks, design, filtering and redesign are all blocked at once.

The CLI-only path that draws candidates from a sequence uses BLAST+ and a human protein sequence database of its own. That path cannot be used from the screen (see What it cannot do), so its readiness is not in this table.

Model weights

The programs alone are not enough. Deep learning models need their weight files as well.

The AlphaFold2 parameters are downloaded automatically on the first run. So the very first time, there is extra download time before design starts. Once they are there, they are not fetched again.

The MSA used when predicting a structure from a sequence alone is not a weight file — it is fetched from an external ColabFold server each time. The same sequence is cached, so the second time is fast.

GPU requirements

Design (BindCraft) and MD (GROMACS) need a GPU. Target lookup, hotspot identification and the report run on CPU.

• Design puts one attempt on one GPU and spreads them across several idle GPUs at once. The current setting lets one job use at most 3
• If no GPU is free, it waits. There is no queue, so several jobs compete with each other
• If the receptor is longer than 450 residues, a window around the hotspots is cut out because of GPU memory

What can run end to end right now

Every step of the on-screen path — structure prediction → hotspots → design → filtering → MD and binding free energy → redesign → report — is prepared. This holds whether you upload a structure or give only a sequence.

What blocks you is elsewhere: per-epitope design has a known defect (see Where people get stuck), and sequence mode and choosing the structure prediction engine are not on the screen (see What it cannot do).

What to do when you are blocked

You cannot install anything yourself. The programs and model weights have to be on the compute server, and that server is the administrator's to touch.

1. Check which step failed in the progress panel. The step name is a row in the table above, and the program on that row is the candidate cause
2. Pass the administrator the step name and the program name, exactly as written
3. If the failure is at the design step, it may not be an install problem — getting no designs at all also happens on a hard target. Try a different site
4. If only the PDF is missing at the report step, the results themselves are intact. Take report.md or report.html instead

Where people get stuck

It throws you out and asks you to sign in. Running requires a sign-in. Viewing the screen does not.

Per-epitope design fails. There is a known defect in this path. Do not turn it on for now.

You cannot sign in with your single sign-on account. The code for contextBio account integration is ready but is not switched on yet on the production server. Sign in with this service's own account.

The length range was set wide and it takes forever. The wider the range, the longer the computation takes and the more the candidates scatter. Start narrow.

A job has turned to failed. When the server restarts, the processes of running jobs disappear, and that state is marked as a failure when it is cleaned up.

Last updated · 2026-09-29